Analysis of medium resolution spectra by automated methods – Application to M 55 and ω Centauri

We have employed feedforward neural networks trained on synthetic spectra in the range 3800 to 5600 Å with resolutions of ~2–3 Å to determine metallicities from spectra of about 1000 main-sequence turn-off, subgiant and red giant stars in the globular clusters M 55 and ω Cen. The overall metallicity accuracies are of the order of 0.15 to 0.2 dex. In addition, we tested how well the stellar parameters $\log\,g$ and Teff can be retrieved from such data without additional colour or photometric information. We find overall uncertainties of 0.3 to 0.4 dex for $\log\,g$ and 140 to 190 K for Teff. In order to obtain some measure of uncertainty for the determined values of [Fe/H], $\log\,g$ and Teff, we applied the bootstrap method for the first time to neural networks for this kind of parametrization problem. The distribution of metallicities for stars in ω Cen clearly shows a large spread in agreement with the well known multiple stellar populations in this cluster.